Study On Image Recognition Method Of Embryo Egg Of Virulent Strain

Vaccination plays an important role in the prevention of influenza and is a more economical and effective means of prevention and control than chemical drugs.It is necessary to detect the quality and activity of egg embryos during vaccine preparation,and eliminate the eggs with low activity and damage,so as to avoid contamination of damaged egg embryos during vaccine preparation and reduce the quality of vaccine.With the development of image recognition technology,the traditional manual detection has some disadvantages,and the research of combining the activity detection of virulent strain egg embryo with machine vision is increasing gradually.In this paper,the egg embryo images of different virulent strains in the incubation period were taken as the research object,and the egg embryo images were recognized by the traditional image recognition method and the convolutional neural network based method respectively,the main contents included:(1)Preparation of data sets.The collected egg embryo images of the virus strains were screened and the samples of egg embryo appeared in the images were retained.To solve the problems of insufficient number of samples and imbalance of categories,the sample number was amplified by means of data enhancement by flipping,rotating and color jitter,and the training set and test set were divided proportionally.(2)The difference between viable and inactivated embryos lies in whether the region of interest contains abundant vascular patterns.In the study of traditional image recognition methods,the image is segmented based on matched filtering,and the skeleton curvature features of blood vessels are introduced.The skeleton curvature and area ratio were extracted to describe the shape features,which were fused with texture features based on gray co-occurrence matrix,and the recognition accuracy of different fusion features in egg embryo image test set was investigated.Experimental results show that the method based on fusion features has a higher recognition accuracy than other single features,in which the average accuracy of the method is 2.0%and 11.33%higher than that of only using shape features and texture features,respectively,and the accuracy of deactivated samples is 3%higher.(3)The method of binary classification and multiple classification of virus strain egg embryo image based on convolutional neural network was studied.The experimental results of CaffeNet and VGG-16 were compared,and the network model CaffeNet,which is more suitable for small data set training,was used for further research.In order to improve the performance of the network based on the data set in this paper,a migration fine-tuning strategy was introduced in the model training with data enhancement as an auxiliary method for the over-fitting phenomenon and the decline of the model generalization ability in the network.The accuracy of the test set was improved by using the fine-adjusted CaffeNet model and the fine-adjusted CaffeNet model after image enhancement.The accuracy of the identification of the second-class egg embryo and the five-class egg embryo reached 98.65%and 94.0%respectively for the second-class and five-class egg embryo.